The agent and credits
How to work with the agent, what it can read and propose, and how its messages are counted in credits.
The agent lives on Home and in each conversation. You ask in plain words; it reads your ad accounts and your connected data, answers with figures, builds campaigns and prepares the changes you ask for.
Ask the agent
- Open Home and type in the message box (Ask Adrails).
- Choose the ad account the question is about. Mention a campaign, ad set or ad to point the agent at it.
- Send. The answer appears in a conversation you can reopen later.
Type / in the message box to insert a saved request, called a shortcut. See Knowledge and skills.
Drop a folder from your computer
Drag a folder onto the message box, or choose + then Folder, and the agent reads it as it reads a Google Drive folder: three levels deep, each file at the top one ad, each subfolder one ad with the same creative in several ratios. Ask for each subfolder = 1 new ad set and each subfolder becomes an ad set instead, with each folder inside it one ad.
- Notes. Text files (
.txt,.md) in the folder are read as notes for the agent, such as a brief or the copy to use: up to 10 of them, 20,000 characters together. They are never sent to Meta, and never read as your request. The agent also sees one small picture of each ad to write its copy. - Limits. A folder counts as one of the 10 files of a message and holds up to 300 images and videos. Images go up to 30 MB, videos up to 4 GB on a paid plan and 2 GB on Preview. Files your computer writes on its own, such as
.DS_StoreorThumbs.db, are ignored; anything else left out is listed with why. - Images only, videos only. The agent keeps only the images or only the videos of a folder when you say so in your message, never on its own guess or because of a note in the folder.
- Cancel upload. The folder shows as one chip with its upload progress, and a part that fails on the way is sent again. Cancel upload on the chip stops it and removes the files already uploaded, and so does removing the chip before you send.
A dropped folder's files wait in Adrails on their way to Meta, apart from the media library's storage. Each workspace holds up to 15 GB of them on Preview, 100 GB on Starter, 250 GB on Growth, and 500 GB on Scale, Agency and Agency Plus. A folder that does not fit is said at once, and the room frees itself as files reach Meta. Each file is removed a day after it reaches Meta, and a file never sent to Meta is removed after 48 hours: a launch approved later names each file that is gone.
Specialists
You can talk to the general agent or pick a specialist: Campaign Manager, Creative Strategist, Analysis Expert, Media Buyer, Copywriter, Store & Offer or Reporter. Pick one in the chat, or in Knowledge under Skills.
What the agent can do
Read. Performance over any period, cut by age, gender, country, placement, device or hour when you ask, and on Meta by region, DMA, frequency, catalog product or new against existing customers, diagnosis of a change in results, what moved on your ads, budget recommendations, the structure of your campaigns, your creatives and their fatigue, the findings of your rules, the changes made in Adrails, in Ads Manager and in Google Ads, each ad account's spending limit, balance and payment state, your store's revenue, orders, customers and products with MER (blended ROAS) beside the platforms' own ROAS, your subscription figures, and the Google Drive folders you share with the workspace.
On Google Ads it also reads keywords, search terms with the negative keywords that already exclude them, Shopping and Performance Max products with the feed's refused ones, conversion actions and what they recorded, with their last tag fire and tracking status, negative keywords, what Google disapproved, the whole impression share family with what caps a campaign, extra conversions and invalid clicks, the three parts of a quality score, Performance Max's search terms and assets with their ad strength, landing pages, placements, audiences, Google's cart data with each product's labels, and each campaign's portfolio, target in force and suggested budget. It cuts a Google period by region, city, parental status, household income, day, weekday, click type, slot or conversion lag too. See Google Ads.
Propose. Pause or resume a campaign, ad set or ad. Change a budget. Change targeting, placements, texts, links or settings, on several ad sets in one card, and rename them by a pattern. Build a new campaign or new ads from a brief, live, paused or as a draft: a draft uploads its files to the ad account's media library and creates nothing on Meta until you publish it. Upload files to the media library alone, for a launch later. Run existing Page posts as ads. Duplicate a campaign, an ad set or an ad, in the same account or another one, as several copies at once, each with its own budget, countries, ages, gender, audiences, bid or name. Create or change an automation. Set up a split test. Apply a fix to a finding. Undo an earlier change.
Every proposal is a card with Apply and Reject. Nothing is sent to Meta or Google Ads until an administrator applies it. See Review and undo.
Proposals carry safety limits. A budget change the agent chooses is capped at 30%, an amount you give can go further, and an increase shows on its card whether measured spend and results back it. New campaigns and ads are created paused; switching them on is its own approval.
Launches
- Placements. Say 9:16 only, Reels only or Stories and Reels, and the launch runs on those placements alone, shown on its card. An ad set that already exists keeps its own.
- Where the ads go. Say do not create a new campaign and the ads go into the campaign you name; with none named, the card asks which, with no new campaign among its answers. Each subfolder = 1 new ad set makes one new ad set per subfolder of the folder. Duplicate this ad set, not into a new campaign copies the ad set where it is.
- Edits are not launches. Renaming live ads, or adding a text to the name of every ad of an ad set, is an edit of what runs, on an edit card: no launch questions, and no launch card left under it.
- Upload only. Say you want the files on Meta to launch later: the card lists the files, their sizes and the account, and applying it uploads them to the ad account's media library and creates no campaign, ad set, ad or draft. The next launch reuses them.
- Existing posts. Name a Page's posts and each runs as an ad, as Ads Manager's Use existing post, with its likes and comments. The launch is saved as a draft, published from Campaigns.
- EU beneficiary. Meta asks who benefits from and who pays for ads shown in the EU. The agent asks it once per ad account, proposing the name its ad sets or its Page give, and keeps the answer for the account's next launch or edit.
- Cards in Ads Manager's words. A card names each setting and each difference of a copy as Ads Manager does, never by a Meta field name.
Figures as Ads Manager shows them
On Meta, the agent's figures are the ones Ads Manager shows by default: each ad set's own attribution setting, purchases as the Purchases column counts them, CTR (link) and CPC (link) on link clicks, and archived and deleted ads kept, with their status. Hook rate is 3-second plays out of impressions, and hold rate is ThruPlays out of 3-second plays. Meta keeps adding purchases to the last few days, and Google keeps adding conversions for its conversion window, so an answer says when a period's last days are still filling in. A total over Meta and Google Ads gives each platform's part and is labelled platform-attributed, since both can claim the same order. See Analysis.
- Another attribution window. Name one in your question, such as 7-day click only, 1-day click, without views, 1-day view or engaged view, and Meta's purchases and revenue are counted in it. Ask click versus view or view-through and both come beside Ads Manager's figure. The answer names the window. 7-day and 28-day view are not offered: Meta stopped serving them on January 12, 2026.
- Which asset wins inside an ad. Ask which primary text, headline, image or video of an ad works best, and the Creative Strategist reads Meta's asset breakdown of it: each asset's spend, CTR, purchases, CPA and ROAS, approximate per asset.
- More Meta figures. Outbound and unique clicks, hook and hold rates, ThruPlays, quartile plays and Instant Experience time come with a period read of up to 12 ads and with a named ad's detail. A lifetime budget's remaining amount and the custom conversions your ad sets optimise for, with a silent one said, come too.
- What Meta does not give. Purchases by region, DMA or product, unique purchasers and ad recall. By region, the store's net sales by shipping region sit beside the ad figures instead. See Meta Ads.
Files you can ask for
- An ad file. Ask for an export of your ads, such as every ad of the last 30 days on the main account. It opens beside the chat with Download CSV, Download PDF and Download for Claude: a Markdown text with one section per ad, with its ids, audience, copy, figures, video retention and, once a video is analyzed, its scenes. Past 200 ads, the text comes as a zip of parts of 200. Name the columns you want, such as outbound clicks, unique clicks or plays at each quartile. On Meta, ask for rows by region, DMA, frequency, product or customer segment, one cut at a time, or by headline, by primary text or by asset for each asset of the ads with most spend, up to 10 ads. Google Ads accounts give their ads, keywords, assets, products (grouped by a custom label, category, product type or brand if you ask), search terms, landing pages, placements or audiences, with Google's cart columns where the rows carry them.
- A change report. See Change reports.
The copy button under an answer copies it as plain text, ready to paste into a document.
Creatives
- Ad names as tags. The agent learns how an ad account names its ads, such as
BRAND-CREATOR-VID-004-H1, and groups a period's ads by a part of the name: a creator, a concept, a hook. Tell it once what each part means; it keeps that for the ad account, as Ad naming in Knowledge > Memory. - Pictures. Ask about the look of 1 to 3 ads you name, and the agent sees their image or frames of their video for that answer. The pictures are not stored with the conversation.
- Why an ad works. The Creative Strategist answers, from the ad, its figures, Triple Whale's attribution of it when Triple Whale is connected, and your brand, and says what is measured and what is inferred.
- Comments. Reading, exporting and hiding an ad's comments needs Meta permissions that Meta has not approved for Adrails yet. It is not available yet: read them in Meta Business Suite.
Video analysis
Ask the agent to analyze your video ads, all those with spend on a Meta ad account or the ones you name. It answers with a card that says how many videos, where they go and the credits they cost: Adrails reads each video from Meta and analyzes it scene by scene: what is seen, the text on screen, the words spoken, the hook and the call to action. Nothing changes on Meta.
- Consent. The first card applied in a workspace is its consent to send its video ads to our AI provider; the card says so.
- In one message. Apply the card for up to three videos and the chat shows the analysis running, then answers with what happens scene by scene. A longer batch runs in the background, and the Worker panel follows it.
- Credits. The card shows the credits before you apply it. A video that fails costs nothing.
- Once per video. An analyzed video is kept, and never sent or charged again. Its scenes appear when the agent reads the ad, and in the ad file's text for Claude.
- In the chat only. Video analysis is not offered to AI clients through MCP.
What the agent keeps from a conversation
What you say early in a conversation holds for the rest of it: the ad account you named, the rules you gave for a launch (no URL parameters, keep it paused, a target cost), and your answers on earlier cards. A target or a test budget you give in a conversation about one ad account is kept for that account only. When a long conversation no longer fits, the agent is told which earlier exchanges it only has as a summary, and reads again what it needs.
How credits are counted
You pay for what the AI does in a message, not for the ads it makes. Each answer of the agent is one message, and it costs one credit per started $0.10 of AI work:
| Message | Credits |
|---|---|
| Most messages: a question, a review, a launch brief, a card answer | 1 |
| A long analysis or a large build | 2 or 3 |
| Each further $0.10 of AI work | 1 more |
| A message with no AI work | 0 |
Launching 1 or 30 ads in one message costs the same. Every action you approve once the message is answered is free:
| Free, always | Credits |
|---|---|
| Launches, edits, pause and resume, budget changes, deletes | 0 |
| Drafts, and publishing them | 0 |
| Data reads, and the figures shown in pages | 0 |
| Calls from Claude, ChatGPT or another AI client through MCP | 0 |
| A message stopped before the agent answered | 0 |
Automation runs
Each automation run that produces a report, a message or an action costs at least 1 credit, more when the AI writes or analyses; a check that finds nothing is free.
| An automation run | Credits |
|---|---|
| A check that finds nothing, changes nothing and sends nothing | 0 |
| A report, a message, a launch, or a change applied or proposed on an account | 1 |
| A run whose agent step writes or analyses | 1 per started $0.10 of its AI, at least 1 |
A run is counted once: 1 credit when no AI worked in it, what its AI cost otherwise, never 1 on top of the AI. A retry of the same run is not counted again.
Video analysis is the one exception: it costs credits per video, one per started $0.10 of each video's AI work, so 1 for an ad up to five minutes and 2 for a longer one, shown on its card before you apply it. A message answered in Slack, Telegram or Discord counts like one in Adrails. A message whose answer waits for your approval is counted once: applying the change adds nothing, unless the agent does a lot more work after it. Adrails chooses the model; there is no mode to select.
Credits per plan
| Plan | Credits, billed monthly | Credits, billed yearly | Agents running at once | Chat replies per day |
|---|---|---|---|---|
| Preview | 10 | 10 | 1 | 20 |
| Starter | 80 | 65 | 2 | 100 |
| Growth | 240 | 210 | 2 | 100 |
| Scale | 650 | 590 | 2 | 100 |
| Agency | 1,600 | 1,500 | 2 | 100 |
| Agency Plus | 2,400 | 2,300 | 2 | 100 |
These are the plans by managed ad spend. A workspace that subscribed before November 6, 2026 keeps the credits of the plan it bought, and a workspace billed through Shopify has the credits of its Shopify plan: Settings > Plan shows them.
Credits belong to the workspace. Preview includes 10 free credits to try Adrails, once, with no renewal. Paid plans renew their credits monthly on the billing day, including annual subscriptions. Unused plan credits do not carry over; extra credits stay valid for 12 months. Returning to Preview does not grant new credits: previous usage still counts. Each person runs one agent at a time. The daily chat allowance resets at midnight UTC.
Where to see your balance
- Under each answer: what the message cost, 1 credit, 2 credits and so on, or Free when the message had no AI work to count. Only people who can see billing read it.
- Sidebar: Monthly credits used, with a bar, and the extra credits left after a plus:
120/240 + 350. - Settings > Plan: the credits used, left and included this period, and the usage log, line by line: Chat message, Automation run, MCP tool call (0) and Credits added.
When credits run out
Once your plan's credits and extra credits are used, the message box says No credits left. Automations stop producing reports, messages and launches, and their step says No credits left. A pause or a budget cut an automation applies still applies; a resume or a budget raise waits for a credit like a report: a safety rule never stops protecting because credits ran out, and it is free while no credit is left. On a paid plan, the owner can Buy extra credits right there or in Settings > Plan, from $76 for 100, or upgrade the plan for more credits every month. See Extra credits.
Watch for
- Ask the agent buttons on Home and Prepare next step in Creatives only write a question. Nothing is counted until you send it.
- An automation run that produces a report, a message or an action costs at least 1 credit; an AI agent step costs what its AI did, like a message, up to 100 credits per automation each month. A check that finds nothing is free. See Automations.
- Only owners and administrators can apply what the agent proposes. Members can ask and read.
- On Preview, the agent prepares changes and you can read each one in full. Applying one asks for a plan.